feat(planner): compose exact functions with summaries

[zzylol]

Closes #171

Why

The planner could build nested accumulator summaries, but it could not preserve a useful inner summary when an exact function had to run across the summary maintenance/read boundary. Those cases collapsed to an opaque raw fallback.

What

• add one general ValueOperation IR node whose semantic operation is independent from ExecutionTiming (MaintenanceTime or ReadTime)
• retain concrete function identity; rate, irate, and increase are distinct intents and irate is no longer lowered as rate
• register accuracy behavior by function definition: exact sum, average, and extrema have separate proven rules; counter rate functions explicitly have no distribution-free approximate-input bound
• preserve unknown accuracy when no applicable rule exists instead of inventing Lipschitz(1); later accuracy-target checks therefore fail closed
• allow runtimes to gate support by concrete function and execution placement
• add typed execution-data-state validation, cost-based global composition selection, CSE, DAG export, and lifecycle traversal

How

Before: an exact fold over a summary readout, or a summary over an exact function without accumulator state, remained an opaque KeepPreAsap region. irate also lost its identity by lowering to Rate, and unclassified exact functions could inherit a placeholder accuracy rule.

After: the planner selects and materializes ValueOperation(operation, timing) around the independently selected summary plan. Function definitions supply semantic identity and accuracy/capability rules; execution timing remains an independent physical property. No function- or issue-specific physical node is introduced.

For exact samples, rate, irate, and increase remain exact. For approximate samples, reset detection and boundary extrapolation do not receive a guessed bound: the guarantee remains unknown unless a deployment registers a proof-backed rule with sufficient evidence.

Review fixes

• Validate concrete operation/child pairs against root targets, instead of rejecting all compositions.
• Retain the caller-model-validated plan through selection and materialization; never recompute its guarantee with the default model.
• Share built-in accuracy and accumulator facts in one function registry. Runtime capability remains deployment-specific.
• Regression tests cover custom-rule acceptance/provenance and rejection of unproven exactness.

Validation

• cargo fmt --all -- --check
• cargo check --workspace
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test -p asap-types --lib (182 passed)
• cargo test -p asap-aware-mapping --lib (392 passed)
• cargo test -p asap-frontend-promql
• cargo test -p asap-integration-tests --test exact_composition (13 passed)
• cargo test --workspace previously reached the environment linker limit for several large DataFusion test binaries (lld bus error); focused suites above pass.